Synthetic Speech, Real Signal: Paralinguistic Preservation and Cross-Lingual Augmentation via Voice Cloning

Alexandra Livia Georgescu; Cole Looney; George Fairs; Luis Miguel San Martin Fernandez; Owen Parsons; Roseline Polle; Stefano Goria; Xiaoliang Wu

REVIEW 2 major objections 5 minor 35 references

Synthetic Speech, Real Signal: Paralinguistic Preservation and Cross-Lingual Augmentation via Voice Cloning

T0 review · 2 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Voice-cloned speech retains enough paralinguistic signal for cross-lingual clinical training data.

desk verdict Sound RQ1 benchmark but the cross-lingual claim overreaches: without a plain Japanese TTS control, you can't tell cloning from language-matched synthesis. read the letter →

arxiv 2607.22304 v1 pith:4AUSYBMC submitted 2026-07-24 cs.LG cs.SD

Roseline Polle , Owen Parsons , George Fairs , Luis Miguel San Martin Fernandez , Cole Looney , Xiaoliang Wu , Alexandra Livia Georgescu , Stefano Goria This is my paper · ORCID

classification cs.LGcs.SD

keywords voicecloningparalinguisticpreservationcross-lingualtransferdepressiondetectionanxietyclinicalspeechdataaugmentationself-supervisedembeddings

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that voice-cloned speech retains enough paralinguistic signal to be used as training data for clinical and affective tasks, including across languages. It benchmarks eight cloning models on five tasks (emotion, sentiment, sarcasm, accent, depression/anxiety) and reports that under the repeat condition the best models keep more than 90% of the above-chance signal, with modest median degradation. Its key cross-lingual claim is that training on English clinical speech cloned into Japanese outperforms raw cross-lingual transfer on real Japanese speech for depression and anxiety detection. A sympathetic reader would care because labelled clinical speech is scarce and mostly English; if cloning preserves the diagnostic signal, synthetic clones could augment low-resource languages and reduce data collection cost.

What carries the argument

The argument is carried by the preservation score P = (A_c − 0.5)/(A_r − 0.5), the fraction of above-chance AUC retained when cloned speech replaces real speech, and by two controlled cloning conditions: repeat (same transcript) and standard (fixed passage), which separate linguistic from paralinguistic content. All measurements go through one probe: WavLM Large embeddings and L2-regularized logistic regression, with AUC as the performance metric. For cross-lingual transfer, English transcripts are machine-translated to Japanese and the original speaker's voice is cloned reading the translation; the resulting synthetic Japanese audio is used as training data and tested on real Japanese recor

What would settle it

Run the cross-lingual experiment again with a different feature probe (e.g., prosodic/spectral features or a different self-supervised model) and a non-linear classifier; if cloned EN→JP no longer beats the raw Real EN baseline on the same real Japanese test set, the headline finding is an artifact of the WavLM + logistic-regression pipeline rather than a property of cloned speech.

Watch

Extended reading notes

Core claim

This paper sets out to show that voice-cloned speech carries enough paralinguistic information to be useful for clinical and affective computing tasks. In its benchmark, under the repeat condition (cloned speech reproduces the original transcript), the best of eight open-source cloning models retain over 90% of above-chance signal measured by the preservation score P, with median degradation of 3.2 percentage points across all configurations. The paper's central cross-lingual finding is that training a depression/anxiety classifier on English clinical speech cloned into Japanese significantly outperforms training on raw English speech when both are tested on real Japanese clinical speech (fo

Load-bearing premise

All results pass through one fixed measurement pipeline—WavLM Large embeddings with L2-regularized logistic regression—so the preservation and transfer numbers could reflect the probe's sensitivity rather than the cloned audio itself.

Editorial extensions

If this is right

Top cloning models retain 90%+ of above-chance paralinguistic signal under repeat condition, so cloned audio can stand in for real data in monolingual tasks with modest degradation.
Cloning English clinical speech into Japanese trains detectors that beat raw cross-lingual transfer on real Japanese test speech, making voice cloning a practical augmentation strategy for low-resource clinical languages.
The repeat-vs-standard comparison shows that much of the preserved signal is carried by linguistic content; paragraph reading, which is content-fixed, still transfers (+2.3pp depression, +4.5pp anxiety), implying acoustic adaptation contributes independently.
Cross-lingual gains appear when training data reaches about 1,000 source speakers and plateau later, so augmentation is feasible at modest scale.
Speaker embedding cosine similarity predicts preservation on clean recordings (r≈0.8–0.9) but fails on noisy ones, so it cannot replace downstream validation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A broader implication not tested here: if the acoustic-adaptation explanation is right, the same cloning pipeline could augment under-represented dialects, ages, or clinical subgroups within a language by cloning a source group into a target group's voice.
The single-probe design suggests the preservation percentages are likely probe-dependent; a richer classifier or different embeddings could show either higher or lower preservation, and the claimed cross-lingual benefit needs replication outside WavLM/LogReg.
The paragraph-translation result hints that machine translation quality matters: if translation introduces errors, the cloned speech inherits them, so gains might grow further with higher-quality translation or back-translation filtering.
The authors' recommendation to validate on a held-out subset before scaling implies a practical test: measure P on the target task at small scale before committing to large synthetic datasets.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

This one is worth reading if you care about synthetic speech for clinical paralinguistic tasks. The main empirical contribution is a clean benchmark: eight open-source voice cloning models, five tasks (emotion, sentiment, sarcasm, accent, depression/anxiety), public plus a large proprietary clinical corpus. They use stratified group 5-fold CV with no speaker leakage, permutation tests against chance, paired bootstrap for cloned vs. real, and a preservation score P that measures fraction of above-chance signal retained. That part is well executed. Under the repeat condition, most models keep 80–90% of the signal, and the best ones clear 90%. That is a genuinely useful data point for anyone thinking about augmenting low-resource clinical speech with cloned audio.

RQ2 is the flashier result: English clinical speech cloned into Japanese beats raw English-to-Japanese transfer by 3–4pp AUC for depression and anxiety on real Japanese speech, and the scaling curves look plausible. But the stress-test is right: the baseline is Real EN audio, so the comparison confounds language/acoustic adaptation with voice cloning. Any Japanese TTS that doesn't clone the source speaker would close much of the language gap, so you cannot attribute the gain to preservation via cloning. The paper says the paragraph-only result shows the gain is "not coming only from translating semantic content," but that just rules out semantic content, not the acoustic language match. This is a moderate weakness, not a fatal one: as a practical augmentation result, cloned data still beats the obvious baseline. But the framing claims more than the experiment shows. They need either a non-cloning Japanese TTS arm or a careful rephrasing.

The other soft spots are minor and mostly acknowledged: all numbers go through a single WavLM Large + LR (C=0.001) probe, so the preservation score is probe-relative; the clinical and RQ2 data are proprietary; only one language pair; the four RQ2 models were selected post-hoc from RQ1 strength. None of these sink RQ1. The limitations section is honest about them. No code or features are released, which limits reproducibility but does not otherwise undermine the experimental design.

For a peer reviewer: the RQ1 benchmark deserves a serious referee, and the paper would be stronger if the RQ2 conclusion were softened or the TTS control added. I'd send it to review.

Referee Report

2 major / 5 minor

Summary. The paper addresses whether voice-cloned synthetic speech preserves paralinguistic signal for downstream classification. It benchmarks eight open-source voice cloning models on five tasks (emotion, sentiment, sarcasm, accent, and depression/anxiety) using public and proprietary clinical datasets. Under a "repeat" condition, most models retain the majority of above-chance discriminative signal (median P=0.87); under a "standard" condition that removes linguistic content, preservation drops substantially. The authors then clone English clinical speech into Japanese and report that training on cloned Japanese speech outperforms training on real English speech when tested on real Japanese speech for depression and anxiety detection. A final analysis correlates speaker embedding similarity with preservation. The central claim is that voice-cloned speech can serve as practical cross-lingual training data for clinical paralinguistic detectors.

Significance. If the findings hold, this is a valuable contribution: it is one of the first systematic evaluations of voice cloning for downstream paralinguistic tasks, covering multiple architectures, tasks, and clinical datasets. The experimental protocol is careful in several respects — stratified group 5-fold CV with no speaker leakage, permutation and paired bootstrap tests, a repeat/standard disentanglement, and a scaling analysis. The paper also identifies an important gap between standard cloning metrics (speaker similarity) and actual downstream preservation. However, the two main empirical claims — monolingual preservation and cross-lingual augmentation — rest on a single fixed feature/classifier probe, and the cross-lingual comparison has a language/synthesis confound that is not resolved by the paragraph-only analysis. These issues are fixable but need attention before the central claims can be accepted.

major comments (2)

[§3.2, Table 4] The cross-lingual augmentation claim (RQ2) compares training on EN→JP cloned speech against training on real English audio (Real EN). These conditions differ simultaneously in language/acoustic space and in synthesis. Any Japanese TTS — even one that discards the source speaker's paralinguistic prosody — would reduce the language mismatch and could improve AUC on Japanese test speech. The paragraph-only result does not remove this confound: the fixed English passage is still translated to Japanese, so the acoustic language changes. The statement that 'the advantage is not coming only from translating semantic content' conflates semantic translation with language-matched acoustic generation. A control condition that synthesizes the same translated transcripts with a standard (non-cloning) Japanese TTS is required to attribute the +3–4pp gains to preservation of the source speaker's parali
[§2.4, §3.1, Figure 1] All preservation and transfer numbers are produced by a single fixed pipeline: WavLM Large embeddings plus L2-regularized logistic regression (C=0.001). The preservation score P is defined relative to the same probe's real-speech AUC. If WavLM embeddings do not capture the paralinguistic cues on which the clinical tasks actually depend, or if the strong regularizer suppresses them, then the conclusion that cloning preserves signal could be an artifact of the probe rather than a property of the cloned audio. The paper acknowledges this in Limitations but does not provide any sensitivity check. I recommend validating the main RQ1/RQ2 conclusions with at least one alternative feature extractor or classifier family (e.g., a larger model fine-tuned on the task, or a spectral/prosodic feature set), or substantively justifying why WavLM+LR suffices for all five tasks.

minor comments (5)

[§3.4] The RQ3 correlation between speaker similarity and AUC degradation is computed on the same WavLM embeddings that feed the downstream classifier. This shared representation may inflate the correlation. The discussion presents RQ3 as a proxy-seeking analysis, so this is not fatal, but the text should note the potential circularity and interpret the correlation as representation-internal rather than fully independent.
[§3.2, Figure 2] The scaling analysis in Figure 2 is only shown for depression; the text says anxiety 'follows a similar pattern' but no figure or table is provided. Since anxiety is a headline result, either include the anxiety scaling plot or state that it is omitted for space and summarize key numbers.
[Table 4] The row label 'Real EN→JP Real' is confusing; 'Real EN' or 'Real EN→JP' would be clearer. Also the text uses 'paired z-test' — please clarify what was paired and whether the test accounts for the 5 bootstrap resamples per seed.
[§3.1] The claim 'All 176 model-task configurations achieve significantly above-chance performance' is not directly verifiable from the tables because the number of configurations is not broken down by task/condition. Please include a supplementary table or clarify the counting.
[§1] Several instances of 'V oice' and 'T able' are spacing artifacts from the source; these should be corrected in the camera-ready version.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the paper is an empirical benchmark, not a fitted derivation; the only minor self-citation is not load-bearing.

full rationale

The paper's claims are empirical measurements rather than a derivation chain. RQ1 compares real versus cloned AUC on public and proprietary datasets under repeat and standard text conditions; RQ2 compares training conditions (JP Real, Real EN, EN-to-JP) on held-out real Japanese test data with bootstrapped significance tests; RQ3 correlates independently computed speaker similarity with AUC degradation. The preservation score P=(Ac-0.5)/(Ar-0.5) is an explicit normalization of measured AUCs, not a fitted parameter renamed as a prediction, and no equation reduces the target claim to its own input. The proprietary corpus is cited to [5], whose authors overlap with the present paper, but the dataset is described with collection details and the central benchmark is also supported by public datasets, so this self-citation is not load-bearing. The RQ2 cross-lingual comparison lacks a non-cloning Japanese TTS control, which is a confound, but a confound is a validity threat rather than a circularity; the paper itself acknowledges the minimal baseline in its Limitations. Thus the appropriate finding is no significant circularity, with only a minor, non-load-bearing self-citation. Score 1 reflects that minor self-citation rather than any actual circular step.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

The central claims rest on a small set of domain assumptions about the measurement pipeline and data matching; no new theoretical entities or fitted constants are introduced. The most load-bearing is the WavLM+LR probe sufficiency, since every preservation and transfer number is filtered through it.

free parameters (1)

Logistic regression regularization C = 0.001
Fixed L2 regularization, not tuned per task; applied identically to real and cloned conditions, so it may compress absolute AUC but should not bias the relative preservation signal.

assumptions (5)

domain assumption WavLM embeddings plus logistic regression is a sufficient probe for the paralinguistic signals relevant to the five tasks.
Section 2.4: all preservation and transfer results are measured through this fixed feature/classifier pipeline; if a task relies on prosodic detail not captured by WavLM or destroyed by the strong regularizer, preservation could be misestimated.
domain assumption In the standard condition, having all speakers produce the same fixed passage removes linguistic content and isolates paralinguistic signal.
Section 2.3: this assumption validates the repeat-vs-standard comparison and the use of paragraph performance as content-independent signal in RQ2.
domain assumption The English and Japanese clinical corpora are matched in protocol, making cross-lingual transfer comparisons meaningful.
Section 2.2: JP was 'collected under the same protocol' but on a different platform; any protocol drift (e.g., recording device, task administration) confounds the RQ2 comparison.
domain assumption Qwen3 235B translation preserves the meaning of the clinical prompts well enough for cross-lingual cloning.
Section 2.3: translated transcripts are the input text for the Japanese cloned speech; translation errors would add noise or shift content, and no quality check is reported.
domain assumption PHQ-8/GAD-7 ≥ 10 plus self-reported prior diagnosis defines the depression/anxiety ground truth.
Section 2.2: the threshold is standard, but labels rely on self-report questionnaires and self-reported history, which are noisy proxies for clinical diagnosis.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Synthetic Speech, Real Signal: Paralinguistic Preservation and Cross-Lingual Augmentation via Voice Cloning." pith.science (2026). https://pith.science/paper/4AUSYBMC

@misc{pith2026260722304,
  author       = {Pith},
  title        = {Pith review of: Synthetic Speech, Real Signal: Paralinguistic Preservation and Cross-Lingual Augmentation via Voice Cloning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4AUSYBMC}},
  note         = {Machine review of arXiv:2607.22304}
}

read the original abstract

Synthetic data augmentation in speech is common practice for linguistic tasks like ASR, but has seen far less work for paralinguistic ones, especially clinical tasks where labelled data is expensive and some patient groups are underrepresented. Voice cloning is one such augmentation approach, but is typically evaluated on speech intelligibility (WER) or speaker similarity (SS) rather than on downstream performance, and it remains unclear whether these preserve the paralinguistic signal such tasks depend on. We benchmark eight voice cloning models on five paralinguistic tasks across public and clinical datasets, showing most preserve signal with modest degradation. We then clone English clinical speech into Japanese and find that training on cloned data outperforms raw cross-lingual transfer for depression and anxiety detection on real Japanese speech, suggesting voice cloning is a promising direction for augmenting clinical speech data in low-resource languages.

Figures

Figures reproduced from arXiv: 2607.22304 by the authors.

**Figure 1.** AUC on the proprietary clinical datasets (depression and anxiety) under both cloning conditions. Circles = Repeat (right annotations); × = Standard (left annotations). Colours denote speech activities (mood, paragraph); the general activity is omitted as results closely mirror paragraph. Dashed lines show Real AUC baselines [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 3.** [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

**Figure 2.** Scaling curves: AUC vs. number of training speakers for depression detection. Shaded bands show 95% CI. pattern, although with cloned models outperforming the baseline from the start (N=50). The in-language Japanese reference remains an upper bound that cloned models approach but do not reach. 3.4. RQ3: Speaker similarity vs. downstream preservation Results show strong Pearson correlations between mean speaker sim… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 3 linked inside Pith

[1]

Existing efforts focus on emotion recogni- tion [7, 8, 9, 10], but these efforts target a single task and none adopts a voice cloning approach

Introduction Synthetic speech augmentation is well explored for linguistic tasks such as Automatic Speech Recognition (ASR) [1, 2, 3, 4], but is less explored in paralinguistic tasks, such as emotion recognition and clinical biomarker detection [5, 6], where the need is greater due to labelled data being more expensive and harder to obtain. Existing effor...
[2]

Method 2.1. Voice Cloning Models We evaluate eight open-source voice cloning models that cover autoregressive, flow-matching, and hybrid architectures (Ta- ble 1), selected for architectural diversity and reported quality. Table 1:Voice cloning models evaluated. Only the models marked with†are used for RQ2 (cross-lingual). Model Architecture XTTS v2 GPT-2...

arXiv 2026
[3]

RQ1: Paralinguistic Preservation Table 3 and Figure 1 report AUC scores for the public and clinical datasets respectively

Results 3.1. RQ1: Paralinguistic Preservation Table 3 and Figure 1 report AUC scores for the public and clinical datasets respectively. All 176 model–task configura- tions achieve significantly above-chance performance (permu- tation test,p<0.05), confirming that paralinguistic signal sur- vives cloning. While 161 show statistically significant degrada- t...

arXiv
[4]

Our results show that, despite this, mod- ern cloning models retain most of the relevant signal

Discussion and Conclusion V oice cloning is typically optimised for intelligibility and speaker similarity, not for preserving the signal that matters in paralinguistic tasks. Our results show that, despite this, mod- ern cloning models retain most of the relevant signal. Under the repeat condition, the best models preserve over 90% of the above-chance di...
[5]

All experimental results and scientific conclusions are solely the authors’

Use of Generative AI Disclosure Generative AI tools were used to assist with manuscript editing. All experimental results and scientific conclusions are solely the authors’
[6]

Spoken language cor- pora augmentation with domain-specific voice-cloned speech,

M. Czy ˙znikiewicz, Ł. Bondaruk, J. Kubiak, A. Wi ˛ acek, Ł. Degórski, M. Kubis, and P. Skórzewski, “Spoken language cor- pora augmentation with domain-specific voice-cloned speech,” in Proceedings of the 19th Conference on Computer Science and In- telligence Systems (FedCSIS), ser. FedCSIS, vol. 39. Polish In- formation Processing Society, Oct. 2024, p. 579–584

2024
[7]

You do not need more data: Improving end- to-end speech recognition by text-to-speech data augmentation,

A. Laptev, R. Korostik, A. Svischev, A. Andrusenko, I. Meden- nikov, and S. Rybin, “You do not need more data: Improving end- to-end speech recognition by text-to-speech data augmentation,” in2020 13th International Congress on Image and Signal Pro- cessing, BioMedical Engineering and Informatics (CISP-BMEI), 2020, pp. 439–444

2020
[8]

Making more of little data: Improving low-resource automatic speech recognition using data augmentation,

M. Bartelds, N. San, B. McDonnell, D. Jurafsky, and M. Wieling, “Making more of little data: Improving low-resource automatic speech recognition using data augmentation,” inProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), A. Rogers, J. Boyd-Graber, and N. Okazaki, Eds. Toronto, Canada: Associ...

2023

Show all 35 references

[9]

An exhaustive evaluation of tts- and vc-based data augmentation for asr,

S. Ogun, V . Colotte, and E. Vincent, “An exhaustive evaluation of tts- and vc-based data augmentation for asr,” 2025

2025
[10]

A multimodal bayesian network for symptom-level depression and anxiety prediction from voice and speech data,

A. Norbury, G. Fairs, A. L. Georgescu, M. M. Nour, E. Molimpakis, and S. Goria, “A multimodal bayesian network for symptom-level depression and anxiety prediction from voice and speech data,”Scientific Reports, vol. 16, no. 1, p. 5397, Feb. 2026

2026
[11]

Speech analysis for health: Current state-of-the-art and the increasing impact of deep learning,

N. Cummins, A. Baird, and B. W. Schuller, “Speech analysis for health: Current state-of-the-art and the increasing impact of deep learning,”Methods, vol. 151, pp. 41–54, 2018

2018
[12]

Generative emotional ai for speech emotion recognition: The case for synthetic emotional speech augmentation,

S. Latif, A. Shahid, and J. Qadir, “Generative emotional ai for speech emotion recognition: The case for synthetic emotional speech augmentation,”Applied Acoustics, vol. 210, p. 109425, 2023

2023
[13]

Towards improving speech emotion recognition using synthetic data augmentation from emotion conversion,

K. M. Ibrahim, A. Perzo, and S. Leglaive, “Towards improving speech emotion recognition using synthetic data augmentation from emotion conversion,” inICASSP 2024 - 2024 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024, pp. 10 636–10 640

2024
[14]

Data aug- mentation using GANs for speech emotion recognition,

A. Chatziagapi, G. Paraskevopoulos, D. Sgouropoulos, G. Pan- tazopoulos, M. Nikandrou, T. Giannakopouloset al., “Data aug- mentation using GANs for speech emotion recognition,” inProc. Interspeech, 2019, pp. 171–175

2019
[15]

A preliminary study on augmenting speech emotion recognition using a diffusion model,

I. Malik, S. Latif, R. Jurdak, and B. W. Schuller, “A preliminary study on augmenting speech emotion recognition using a diffusion model,”arXiv preprint arXiv:2305.11413, 2023

2023 arXiv
[16]

V oice cloning: Comprehensive sur- vey,

H. Azzuni and A. E. Saddik, “V oice cloning: Comprehensive sur- vey,” 2025

2025
[17]

ClonEval: An open voice cloning benchmark,

I. Christop, T. Kuczy ´nski, and M. Kubis, “ClonEval: An open voice cloning benchmark,” 2025

2025
[18]

The V oicePrivacy 2024 challenge evaluation plan,

N. Tomashenko, X. Miao, P. Champion, S. Meyer, X. Wang, E. Vincentet al., “The V oicePrivacy 2024 challenge evaluation plan,” 2024

2024
[19]

A review of depression and suicide risk assessment using speech analysis,

N. Cummins, S. Scherer, J. Krajewski, S. Schnieder, J. Epps, and T. F. Quatieri, “A review of depression and suicide risk assessment using speech analysis,”Speech Communication, vol. 71, pp. 10– 49, 2015

2015
[20]

Automated assess- ment of psychiatric disorders using speech: A systematic review,

D. M. Low, K. H. Bentley, and S. S. Ghosh, “Automated assess- ment of psychiatric disorders using speech: A systematic review,” Laryngoscope Investigative Otolaryngology, vol. 5, no. 1, pp. 96– 116, 2020

2020
[21]

Kintsugi voice biomarker technology,

Kintsugi Health, “Kintsugi voice biomarker technology,” 2024, accessed: 2025-06-01. [Online]. Available: https://www. kintsugihealth.com

2024
[22]

V ocal biomarker platform,

Sonde Health, “V ocal biomarker platform,” 2024, accessed: 2025-06-01. [Online]. Available: https://www.sondehealth.com

2024
[23]

The Distress Analysis Interview Corpus of human and computer interviews,

J. Gratch, R. Artstein, G. Lucas, G. Stratou, S. Scherer, A. Nazar- ianet al., “The Distress Analysis Interview Corpus of human and computer interviews,” inProceedings of the Ninth Interna- tional Conference on Language Resources and Evaluation (LREC 2014). Reykjavik, Iceland:...

2014
[24]

The Androids corpus: A new publicly available benchmark for speech based depression detection,

F. Tao, A. Esposito, and A. Vinciarelli, “The Androids corpus: A new publicly available benchmark for speech based depression detection,” inProc. Interspeech, 2023, pp. 4149–4153

2023
[25]

DEPAC: a corpus for depression and anxiety detection from speech,

M. Tasnim, M. Ehghaghi, B. Diep, and J. Novikova, “DEPAC: a corpus for depression and anxiety detection from speech,” inPro- ceedings of the Eighth Workshop on Computational Linguistics and Clinical Psychology. Seattle, USA: Association for Compu- tational Linguistics, 2022, pp. 1–16

2022
[26]

IEMOCAP: Interactive emotional dyadic motion capture database,

C. Busso, M. Bulut, C.-C. Lee, A. Kazemzadeh, E. Mower, S. Kimet al., “IEMOCAP: Interactive emotional dyadic motion capture database,”Language Resources and Evaluation, vol. 42, no. 4, pp. 335–359, 2008

2008
[27]

MELD: A multimodal multi-party dataset for emo- tion recognition in conversations,

S. Poria, D. Hazarika, N. Majumder, G. Naik, E. Cambria, and R. Mihalcea, “MELD: A multimodal multi-party dataset for emo- tion recognition in conversations,” inProceedings of the 57th An- nual Meeting of the Association for Computational Linguistics, 2019, pp. 527–536

2019
[28]

Towards multimodal sarcasm detection (an obviously perfect paper),

S. Castro, D. Hazarika, V . Pérez-Rosas, R. Zimmermann, R. Mi- halcea, and S. Poria, “Towards multimodal sarcasm detection (an obviously perfect paper),” inProceedings of the 57th Annual Meeting of the Association for Computational Linguistics, 2019, pp. 4619–4629

2019
[29]

CSTR VCTK cor- pus: English multi-speaker corpus for CSTR voice cloning toolkit (version 0.92),

J. Yamagishi, C. Veaux, and K. MacDonald, “CSTR VCTK cor- pus: English multi-speaker corpus for CSTR voice cloning toolkit (version 0.92),” 2019

2019
[30]

Robust speech recognition via large-scale weak su- pervision,

A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak su- pervision,” inInternational Conference on Machine Learning. PMLR, 2023, pp. 28 492–28 518

2023
[31]

Qwen3 technical report,

A. Yang, B. Yang, B. Zhanget al., “Qwen3 technical report,” arXiv preprint arXiv:2505.09388, 2025

2025 arXiv
[32]

WavLM: Large-scale self-supervised pre-training for full stack speech processing,

S. Chen, C. Wang, Z. Chen, Y . Wu, S. Liu, Z. Chenet al., “WavLM: Large-scale self-supervised pre-training for full stack speech processing,”IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 6, pp. 1505–1518, 2022

2022
[33]

Paral- bench: A large-scale benchmark for computational paralinguistics over acoustic foundation models,

Z. Zhang, W. Xu, Z. Dong, K. Wang, Y . Wu, J. Penget al., “Paral- bench: A large-scale benchmark for computational paralinguistics over acoustic foundation models,”IEEE Transactions on Affective Computing, vol. 16, no. 3, pp. 1290–1306, 2025

2025
[34]

Anxiety and depression as bidirectional risk factors for one another: A meta-analysis of lon- gitudinal studies,

N. C. Jacobson and M. G. Newman, “Anxiety and depression as bidirectional risk factors for one another: A meta-analysis of lon- gitudinal studies,”Psychological bulletin, vol. 143, no. 11, Nov. 2017

2017
[35]

Comor- bidity within mental disorders: a comprehensive analysis based on 145 990 survey respondents from 27 countries,

J. J. McGrath, C. C. W. Lim, O. Plana-Ripollet al., “Comor- bidity within mental disorders: a comprehensive analysis based on 145 990 survey respondents from 27 countries,”Epidemiology and Psychiatric Sciences, vol. 29, p. e153, Aug. 2020

2020

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.

[1] [1]

Existing efforts focus on emotion recogni- tion [7, 8, 9, 10], but these efforts target a single task and none adopts a voice cloning approach

Introduction Synthetic speech augmentation is well explored for linguistic tasks such as Automatic Speech Recognition (ASR) [1, 2, 3, 4], but is less explored in paralinguistic tasks, such as emotion recognition and clinical biomarker detection [5, 6], where the need is greater due to labelled data being more expensive and harder to obtain. Existing effor...

[2] [2]

Method 2.1. Voice Cloning Models We evaluate eight open-source voice cloning models that cover autoregressive, flow-matching, and hybrid architectures (Ta- ble 1), selected for architectural diversity and reported quality. Table 1:Voice cloning models evaluated. Only the models marked with†are used for RQ2 (cross-lingual). Model Architecture XTTS v2 GPT-2...

arXiv 2026

[3] [3]

RQ1: Paralinguistic Preservation Table 3 and Figure 1 report AUC scores for the public and clinical datasets respectively

Results 3.1. RQ1: Paralinguistic Preservation Table 3 and Figure 1 report AUC scores for the public and clinical datasets respectively. All 176 model–task configura- tions achieve significantly above-chance performance (permu- tation test,p<0.05), confirming that paralinguistic signal sur- vives cloning. While 161 show statistically significant degrada- t...

arXiv

[4] [4]

Our results show that, despite this, mod- ern cloning models retain most of the relevant signal

Discussion and Conclusion V oice cloning is typically optimised for intelligibility and speaker similarity, not for preserving the signal that matters in paralinguistic tasks. Our results show that, despite this, mod- ern cloning models retain most of the relevant signal. Under the repeat condition, the best models preserve over 90% of the above-chance di...

[5] [5]

All experimental results and scientific conclusions are solely the authors’

Use of Generative AI Disclosure Generative AI tools were used to assist with manuscript editing. All experimental results and scientific conclusions are solely the authors’

[6] [6]

Spoken language cor- pora augmentation with domain-specific voice-cloned speech,

M. Czy ˙znikiewicz, Ł. Bondaruk, J. Kubiak, A. Wi ˛ acek, Ł. Degórski, M. Kubis, and P. Skórzewski, “Spoken language cor- pora augmentation with domain-specific voice-cloned speech,” in Proceedings of the 19th Conference on Computer Science and In- telligence Systems (FedCSIS), ser. FedCSIS, vol. 39. Polish In- formation Processing Society, Oct. 2024, p. 579–584

2024

[7] [7]

You do not need more data: Improving end- to-end speech recognition by text-to-speech data augmentation,

A. Laptev, R. Korostik, A. Svischev, A. Andrusenko, I. Meden- nikov, and S. Rybin, “You do not need more data: Improving end- to-end speech recognition by text-to-speech data augmentation,” in2020 13th International Congress on Image and Signal Pro- cessing, BioMedical Engineering and Informatics (CISP-BMEI), 2020, pp. 439–444

2020

[8] [8]

Making more of little data: Improving low-resource automatic speech recognition using data augmentation,

M. Bartelds, N. San, B. McDonnell, D. Jurafsky, and M. Wieling, “Making more of little data: Improving low-resource automatic speech recognition using data augmentation,” inProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), A. Rogers, J. Boyd-Graber, and N. Okazaki, Eds. Toronto, Canada: Associ...

2023